Anthropic's in-house version control system Scion was fully disclosed for the first time in an Anthropic engineering post on October 13, 2026: over five months, the company migrated its monorepo off Git entirely onto Scion, while also building and shipping its own code review surface called Antique, plus code search and a source control permission system. What forced the move was not a single outage but a set of numbers spinning out of control: Anthropic averaged around 2,500 code changes merged per day in May 2026; by September, it was merging over 30,000 per day.
What Git could no longer carry was agent-scale traffic
The history of Anthropic's main branch alone is almost 100 GB, and a compressed working copy is over 25 GB. CI performs on the order of 25 million code checkouts per day, before counting developer environments, agent sandboxes and research jobs; fetched naively, that would be over 600 PB of Git traffic a day. The team had already used every standard workaround — shallow clones, sparse checkouts, promisor remotes, several hundred internal Git mirrors, repository snapshots in blob storage — but as the merge rate kept climbing, seeds fell further behind the tip of main, and history-dependent operations such as diff and blame grew harder to keep working on partial clones. Anthropic's conclusion: in an environment where nearly all clients are always-online agents and jobs, the offline, decentralized assumptions that made Git great had become the thing holding it back.
Scion's answer: move the repository back to the server
Scion is designed around principles that invert Git's. Clients are assumed to be always online and fetch data from the server when they need it; the server answers history and blame queries, creates commits and performs merges and rebases, while the client keeps only a bounded local object cache. The working copy becomes optional and sparse by design: a CI job that needs the diff between two commits can simply ask the server, skipping checkout entirely, and jobs that write code can often work through the API as well. Commits are meant to be cheap, lock-free and priced by the size of the change, not the size of the repo: Scion's unit of work, the changeset, records only the files changed relative to its first parent plus content-addressed pointers to file contents, and becomes available everywhere instantly, with no packfiles and nothing to push. In production, Scion's infrastructure footprint is about 200x smaller than the previous Git mirroring setup, checkouts run two orders of magnitude faster, and CI alone realized a 25% CPU capacity win.
The engineer's new job: telling Claude "no"
The part of this postmortem most worth other teams' attention is how the human role changed. A team of seven built both Scion and Antique and completed the migration in five months, while daily merges grew tenfold. The engineers rate Claude as better than themselves at writing code now, but they still supply the high-level understanding of how systems fit together, the design trade-offs and the judgment about required scale. The examples are concrete: agents sometimes violated core system constraints to satisfy a pattern that should never have been allowed, and some systems tried to brute-force merge times by linearly reading the commit log instead of asking the code review tool. Adversarial review is effective but can be myopic, at one point pushing Claude to pile excessive validation onto a component out of distrust of another internal service. So the engineers now review the intent of each change rather than reading every line — a small detail that says more about how agentic programming divides labor in its deep-water phase than any benchmark.
Anthropic states the limits plainly at the end: Scion is not a Git replacement for general use. It only makes sense for truly enormous monorepos with tens of thousands of engineers or swarms of agents writing code prolifically, Anthropic's ability to spend tokens internally far exceeds what most companies would accept, and a broadly usable public version is unlikely. The company shared the design because it expects more heavy users of AI coding to hit the same wall within six to twelve months: the bottleneck moves from writing code to infrastructure built for human scale — source control, builds and CI. For most teams, the realistic move right now is to measure how close their own checkout and merge volumes are to that inflection point, not to rush into building a new system.
